← Back to guides
JevMade field notes / Architecture and workflow guide
system1-agents
An agent-runtime guide that separates decision turns from planning and prose, provides multiple task fronts, and documents replayable comparisons.
Original by ThinkFlowLabEvaluationRepository README and documentationSource reviewed
Before you dive in
What you’ll find in the original
- Place a decision model only in turns with a known option set; leave planning, typed values, and final prose to a chat model.
- Compare agents on the same episode and environment, retaining replays rather than quoting model-call latency alone.
- Use the documented model interface to swap backends without rewriting each browser, game, or routing front.
Worth knowingBehavior and measurements are reported by the repository authors; source and critical implementation were inspected but not executed.